Solving a Sudoku puzzle using computer vision and CUDA.
Final project for University of South Carolina Parallel Computing class, CSCE 569.

• Michael Brunson
• Stephen Still

• OpenCV 2.4.12
• CUDA (Toolkit 7.5, Compute 5.0)

OpenCV is used for extraction of the Sudoku puzzle from a video feed. The main.h file specifies has a define statement that allows the use the GPU for performing the machine vision algorithms including thresholding and edge detection.

After thresholding and filtering, the digit images are extracted and recognized using OpenCV's KNearest class. This uses a K nearest neighbor algorithm with the images in the "train" folder as the models for comparison. The dataset is small and made for printed text and should work with simple computer-generated texts.

The GPU is always used to solve the puzzle although the method is not as efficient as some basic CPU algorithms.

This code was run on a Dell Alienware 15 with a Intel Core i7-4710HQ CPU and Nvidia Geforce GTX 970M GPU. When running this on a different architecture, you will be required to build OpenCV from source and specify the proper architecture for CUDA in CMake (in my case the "CUDA_ARCH_BIN" value had to be "2.0 3.0 3.5 5.0" and the "CUDA_ARCH_PTX" value had to be "5.0" to get OpenCV to interface with the GPU).

Extraction: CPU: ~1200 ms
GPU: ~2100 ms

Solving: CPU: < 2 ms
GPU: ~200 ms

Based on the algorithms we used, the GPU did not speed up the program. Using OpenCV's Mat data structure resulted in too much overhead for the extraction algorithm and some parts of the extraction process are not easily parallelizeable. This resulted in spending too much time copying back and forth from host to GPU.

The solving algorthim suffers from similar issues. The breadth first search algorithm has not been perfected on the GPU yet, so the implementation here is inefficient and results in way too much overhead.

Better algorithms are needed!